Vakyartha documentation

1use futures::StreamExt;
2use serde_json::Value;
3use tokio_util::sync::CancellationToken;
4
5use crate::Provider;
6use crate::error::LlmError;
7use crate::gate::ProviderGate;
8use crate::sse::SseDecoder;
9use crate::stream::{EventStream, StreamEvent, channel};
10use crate::types::{
11 AssistantMessage, ChatRequest, ContentBlock, Message, Role, StopReason, ToolDefinition, Usage,
12};
13
14pub const OPENAI_DEFAULT_BASE_URL: &str = "https://api.openai.com/v1";
15
16#[derive(Debug, Clone, Default)]
17pub struct OpenAiConfig {
18 pub api_key: String,
19 pub base_url: String,
20 /// When true, and the request carries `ChatRequest::cache`, send
21 /// `prompt_cache_key` so the provider can route repeat traffic to the
22 /// same cache-warm backend.
23 pub cache_key: bool,
24 /// When true, also send OpenRouter's `session_id` alongside
25 /// `prompt_cache_key` — OpenRouter accepts both and uses `session_id`
26 /// to pin the upstream that holds the cache.
27 pub openrouter: bool,
28}
29
30/// Batch transcription through the OpenAI-compatible `/audio/transcriptions`
31/// endpoint. The model is supplied by discovery/configuration; this adapter
32/// deliberately has no baked-in model catalogue or default.
33pub async fn transcribe(
34 config: &OpenAiConfig,
35 audio: &[u8],
36 mime: &str,
37 model: &str,
38 cancel: &CancellationToken,
39) -> Result<String, LlmError> {
40 if audio.is_empty() || mime.trim().is_empty() || model.trim().is_empty() {
41 return Err(LlmError::InvalidRequest(
42 "audio, mime, and model are required".into(),
43 ));
44 }
45 if cancel.is_cancelled() {
46 return Err(LlmError::Aborted { partial: None });
47 }
48 let filename = if mime.contains("wav") {
49 "audio.wav"
50 } else if mime.contains("mpeg") || mime.contains("mp3") {
51 "audio.mp3"
52 } else if mime.contains("ogg") {
53 "audio.ogg"
54 } else if mime.contains("oga") {
55 "audio.oga"
56 } else if mime.contains("webm") {
57 "audio.webm"
58 } else if mime.contains("mp4") || mime.contains("m4a") {
59 "audio.m4a"
60 } else if mime.contains("flac") {
61 "audio.flac"
62 } else {
63 "audio.bin"
64 };
65 let part = reqwest::multipart::Part::bytes(audio.to_vec())
66 .file_name(filename)
67 .mime_str(mime)
68 .map_err(|e| LlmError::InvalidRequest(e.to_string()))?;
69 let form = reqwest::multipart::Form::new()
70 .part("file", part)
71 .text("model", model.to_string());
72 let url = format!(
73 "{}/audio/transcriptions",
74 config.base_url.trim_end_matches('/')
75 );
76 let response = tokio::select! {
77 _ = cancel.cancelled() => return Err(LlmError::Aborted { partial: None }),
78 result = reqwest::Client::new().post(url).bearer_auth(&config.api_key).multipart(form).send() => result.map_err(|e| LlmError::Network(e.to_string()))?,
79 };
80 let status = response.status().as_u16();
81 let value: Value = response
82 .json()
83 .await
84 .map_err(|e| LlmError::Parse(e.to_string()))?;
85 if status >= 400 {
86 return Err(LlmError::InvalidRequest(format!(
87 "transcription provider returned HTTP {status}: {value}"
88 )));
89 }
90 let text = value
91 .get("text")
92 .and_then(Value::as_str)
93 .unwrap_or("")
94 .trim()
95 .to_string();
96 Ok(text)
97}
98
99/// Synthesize speech through an OpenAI-compatible `/audio/speech` endpoint.
100/// The model is always explicit and the response is bounded before it is
101/// materialized, so a misbehaving provider cannot exhaust the process.
102pub async fn speak(
103 config: &OpenAiConfig,
104 text: &str,
105 model: &str,
106 voice: Option<&str>,
107 format: &str,
108 cancel: &CancellationToken,
109) -> Result<Vec<u8>, LlmError> {
110 if text.trim().is_empty() || model.trim().is_empty() || format.trim().is_empty() {
111 return Err(LlmError::InvalidRequest(
112 "text, model, and format are required".into(),
113 ));
114 }
115 if cancel.is_cancelled() {
116 return Err(LlmError::Aborted { partial: None });
117 }
118 let mut body = serde_json::json!({
119 "model": model,
120 "input": text,
121 "response_format": format,
122 });
123 if let Some(voice) = voice.filter(|v| !v.trim().is_empty()) {
124 body["voice"] = Value::String(voice.to_string());
125 }
126 let url = format!("{}/audio/speech", config.base_url.trim_end_matches('/'));
127 let response = tokio::select! {
128 _ = cancel.cancelled() => return Err(LlmError::Aborted { partial: None }),
129 result = reqwest::Client::new().post(url).bearer_auth(&config.api_key).json(&body).send() => result.map_err(|e| LlmError::Network(e.to_string()))?,
130 };
131 let status = response.status().as_u16();
132 if status >= 400 {
133 let body = response.text().await.unwrap_or_default();
134 return Err(map_status_error(status, &body, None));
135 }
136 const MAX_AUDIO_BYTES: usize = 16 * 1024 * 1024;
137 if response
138 .content_length()
139 .is_some_and(|n| n > MAX_AUDIO_BYTES as u64)
140 {
141 return Err(LlmError::InvalidRequest(
142 "provider audio exceeds 16 MiB".into(),
143 ));
144 }
145 let mut output = Vec::new();
146 let mut stream = response.bytes_stream();
147 while let Some(chunk) = tokio::select! {
148 _ = cancel.cancelled() => return Err(LlmError::Aborted { partial: None }),
149 chunk = stream.next() => chunk,
150 } {
151 let chunk = chunk.map_err(|e| LlmError::Network(e.to_string()))?;
152 if output.len().saturating_add(chunk.len()) > MAX_AUDIO_BYTES {
153 return Err(LlmError::InvalidRequest(
154 "provider audio exceeds 16 MiB".into(),
155 ));
156 }
157 output.extend_from_slice(&chunk);
158 }
159 if output.is_empty() {
160 return Err(LlmError::Parse("provider returned empty audio".into()));
161 }
162 Ok(output)
163}
164
165#[derive(Clone)]
166pub struct OpenAiCompletionsProvider {
167 http: reqwest::Client,
168 config: OpenAiConfig,
169 gate: ProviderGate,
170}
171
172impl OpenAiCompletionsProvider {
173 pub fn new(config: OpenAiConfig) -> Result<Self, LlmError> {
174 let http = reqwest::Client::builder()
175 .connect_timeout(std::time::Duration::from_secs(30))
176 .build()
177 .map_err(|e| LlmError::Network(e.to_string()))?;
178 Ok(OpenAiCompletionsProvider {
179 http,
180 gate: ProviderGate::new(&config.base_url, &config.api_key),
181 config,
182 })
183 }
184}
185
186pub fn build_body(config: &OpenAiConfig, request: &ChatRequest) -> Result<Value, LlmError> {
187 // Chat Completions carries no reasoning-item channel, so `Thinking`
188 // blocks are dropped unconditionally by `append_message` below — there
189 // is no turn boundary to compute here (contrast Anthropic/Google, which
190 // must replay signed thinking within the current turn).
191 let mut messages: Vec<Value> = Vec::with_capacity(request.messages.len() + 1);
192 if let Some(system) = &request.system {
193 messages.push(serde_json::json!({"role": "system", "content": system}));
194 }
195 for m in &request.messages {
196 append_message(&mut messages, m)?;
197 }
198
199 let mut body = serde_json::json!({
200 "model": request.model,
201 "messages": messages,
202 "stream": true,
203 "stream_options": {"include_usage": true},
204 });
205 if let Some(cache) = &request.cache {
206 if config.cache_key {
207 body["prompt_cache_key"] = serde_json::json!(cache.session_key);
208 }
209 if config.openrouter {
210 body["session_id"] = serde_json::json!(cache.session_key);
211 }
212 }
213 if !request.tools.is_empty() {
214 let tools: Vec<Value> = request
215 .tools
216 .iter()
217 .map(|t: &ToolDefinition| {
218 serde_json::json!({
219 "type": "function",
220 "function": {
221 "name": t.name,
222 "description": t.description,
223 "parameters": t.parameters,
224 },
225 })
226 })
227 .collect();
228 body["tools"] = Value::Array(tools);
229 }
230 Ok(body)
231}
232
233fn append_message(out: &mut Vec<Value>, m: &Message) -> Result<(), LlmError> {
234 match m.role {
235 Role::User => {
236 let mut text_parts: Vec<&str> = Vec::new();
237 let mut image_parts: Vec<String> = Vec::new();
238 let mut tool_results: Vec<&ContentBlock> = Vec::new();
239 for b in &m.content {
240 match b {
241 ContentBlock::Text { text } => text_parts.push(text),
242 ContentBlock::Image { source } => {
243 // Data URLs are the transport-agnostic form for the
244 // chat-completions API.
245 image_parts
246 .push(format!("data:{};base64,{}", source.media_type, source.data));
247 }
248 ContentBlock::ToolResult { .. } => tool_results.push(b),
249 ContentBlock::ToolUse { .. } => {
250 return Err(LlmError::InvalidRequest(
251 "tool_use blocks must appear in assistant messages".into(),
252 ));
253 }
254 ContentBlock::Thinking { .. } => {}
255 // Only the Anthropic adapter understands server-side
256 // tool search; every other adapter skips this opaque
257 // block entirely (docs/design/68 §5/§12).
258 ContentBlock::Provider { .. } => {}
259 }
260 }
261 for r in tool_results {
262 let ContentBlock::ToolResult {
263 tool_use_id,
264 content,
265 ..
266 } = r
267 else {
268 unreachable!()
269 };
270 out.push(serde_json::json!({
271 "role": "tool",
272 "tool_call_id": tool_use_id,
273 "content": content,
274 }));
275 }
276 if !image_parts.is_empty() || !text_parts.is_empty() {
277 let content = if image_parts.is_empty() {
278 serde_json::json!(text_parts.join("\n"))
279 } else {
280 let mut parts: Vec<Value> = text_parts
281 .iter()
282 .map(|t| serde_json::json!({"type": "text", "text": t}))
283 .collect();
284 for url in &image_parts {
285 parts.push(serde_json::json!({
286 "type": "image_url",
287 "image_url": {"url": url}
288 }));
289 }
290 serde_json::json!(parts)
291 };
292 out.push(serde_json::json!({
293 "role": "user",
294 "content": content,
295 }));
296 }
297 }
298 Role::Assistant => {
299 let mut text = String::new();
300 let mut tool_calls: Vec<Value> = Vec::new();
301 for b in &m.content {
302 match b {
303 ContentBlock::Text { text: t } => {
304 if !text.is_empty() {
305 text.push('\n');
306 }
307 text.push_str(t);
308 }
309 ContentBlock::ToolUse { id, name, input } => {
310 tool_calls.push(serde_json::json!({
311 "id": id,
312 "type": "function",
313 "function": {
314 "name": name,
315 "arguments": serde_json::to_string(input)
316 .map_err(|e| LlmError::Parse(e.to_string()))?,
317 },
318 }));
319 }
320 ContentBlock::Thinking { .. }
321 | ContentBlock::ToolResult { .. }
322 | ContentBlock::Image { .. }
323 | ContentBlock::Provider { .. } => {}
324 }
325 }
326 let mut msg = serde_json::json!({"role": "assistant"});
327 if !text.is_empty() || tool_calls.is_empty() {
328 msg["content"] = Value::String(text);
329 } else {
330 msg["content"] = Value::Null;
331 }
332 if !tool_calls.is_empty() {
333 msg["tool_calls"] = Value::Array(tool_calls);
334 }
335 out.push(msg);
336 }
337 }
338 Ok(())
339}
340
341fn map_status_error(status: u16, body: &str, retry_after: Option<u64>) -> LlmError {
342 let message = serde_json::from_str::<Value>(body)
343 .ok()
344 .and_then(|v| {
345 v.pointer("/error/message")
346 .or_else(|| v.pointer("/message"))
347 .and_then(|m| m.as_str().map(String::from))
348 })
349 .unwrap_or_else(|| body.chars().take(500).collect());
350
351 match status {
352 401 | 403 => LlmError::Auth(message),
353 400 => match LlmError::classify_400(message.clone()) {
354 over_length @ LlmError::Context(_) => over_length,
355 _ => LlmError::invalid_request_for_endpoint("/v1/chat/completions", message),
356 },
357 404 | 413 | 422 => LlmError::invalid_request_for_endpoint("/v1/chat/completions", message),
358 429 => LlmError::RateLimit {
359 message,
360 retry_after_secs: retry_after,
361 },
362 503 | 529 => LlmError::Overloaded(message),
363 _ => LlmError::Api { status, message },
364 }
365}
366
367struct Accumulator {
368 message: AssistantMessage,
369 saw_end: bool,
370 tool_pos: std::collections::HashMap<usize, usize>,
371 raw_json: std::collections::HashMap<usize, String>,
372}
373
374impl Accumulator {
375 fn new(model: &str) -> Self {
376 Accumulator {
377 message: AssistantMessage::empty(model),
378 saw_end: false,
379 tool_pos: std::collections::HashMap::new(),
380 raw_json: std::collections::HashMap::new(),
381 }
382 }
383
384 fn convert(&mut self, data: &str) -> Result<Option<StreamEvent>, LlmError> {
385 let v: Value = serde_json::from_str(data)
386 .map_err(|e| LlmError::Parse(format!("bad chunk json: {e}")))?;
387
388 if let Some(usage) = v.get("usage").filter(|u| !u.is_null()) {
389 // OpenAI's `prompt_tokens` is the WHOLE prompt, cache hits
390 // included (`prompt_tokens_details.cached_tokens` is a subset
391 // of it, not an addition to it). Normalized `input_tokens` is
392 // only the non-cached remainder, matching every other adapter
393 // (docs/design/68-context-engine.md §1) — `Usage::prompt_tokens()`
394 // reconstructs the original total.
395 let prompt_tokens = usage
396 .get("prompt_tokens")
397 .and_then(|x| x.as_u64())
398 .unwrap_or(0);
399 let cached_tokens = usage
400 .get("prompt_tokens_details")
401 .and_then(|d| d.get("cached_tokens"))
402 .and_then(|x| x.as_u64())
403 .unwrap_or(0);
404 self.message.usage = Usage {
405 input_tokens: prompt_tokens.saturating_sub(cached_tokens),
406 output_tokens: usage
407 .get("completion_tokens")
408 .and_then(|x| x.as_u64())
409 .unwrap_or(0),
410 cache_read_input_tokens: (cached_tokens > 0).then_some(cached_tokens),
411 cache_creation_input_tokens: None,
412 ..Default::default()
413 };
414 }
415
416 let Some(choice) = v.get("choices").and_then(|c| c.get(0)) else {
417 return Ok(None);
418 };
419
420 #[allow(clippy::dbg_macro)]
421 if std::env::var_os("VAK_LLM_DEBUG").is_some() {
422 eprintln!(
423 "[vak-llm] frame finish={:?} delta_keys={:?} content_len={}",
424 choice.get("finish_reason"),
425 choice.get("delta").map(|d| d
426 .as_object()
427 .map(|o| o.keys().cloned().collect::<Vec<_>>())
428 .unwrap_or_default()),
429 self.message.content.len()
430 );
431 }
432 if let Some(finish) = choice.get("finish_reason").and_then(|f| f.as_str()) {
433 // Accumulated tool_use blocks are ground truth: some compat
434 // endpoints close tool-call turns with finish reasons outside
435 // the canonical set (e.g. plain "stop"). Trusting the label
436 // would strand a dangling tool_use and kill the run.
437 let has_tool_use = self
438 .message
439 .content
440 .iter()
441 .any(|b| matches!(b, ContentBlock::ToolUse { .. }));
442 self.message.stop_reason = if has_tool_use {
443 StopReason::ToolUse
444 } else {
445 match finish {
446 "tool_calls" | "function_call" => StopReason::ToolUse,
447 "length" => StopReason::MaxTokens,
448 _ => StopReason::EndTurn,
449 }
450 };
451 self.saw_end = true;
452 return Ok(Some(StreamEvent::End {
453 message: self.message.clone(),
454 }));
455 }
456
457 let Some(delta) = choice.get("delta") else {
458 return Ok(None);
459 };
460
461 let mut event = None;
462 if let Some(text) = delta
463 .get("content")
464 .and_then(|c| c.as_str())
465 .filter(|t| !t.is_empty())
466 {
467 append_text_block(&mut self.message.content, text);
468 event = Some(StreamEvent::TextDelta {
469 delta: text.to_string(),
470 partial: self.message.clone(),
471 });
472 }
473
474 if let Some(calls) = delta.get("tool_calls").and_then(|c| c.as_array()) {
475 for call in calls {
476 let idx = call.get("index").and_then(|i| i.as_u64()).unwrap_or(0) as usize;
477 let pos = match self.tool_pos.get(&idx) {
478 Some(&p) => p,
479 None => {
480 self.message.content.push(ContentBlock::ToolUse {
481 id: String::new(),
482 name: String::new(),
483 input: Value::Object(Default::default()),
484 });
485 let p = self.message.content.len() - 1;
486 self.tool_pos.insert(idx, p);
487 p
488 }
489 };
490 if let ContentBlock::ToolUse { id, name, .. } = &mut self.message.content[pos] {
491 if let Some(new_id) = call.get("id").and_then(|i| i.as_str()) {
492 *id = new_id.to_string();
493 }
494 if let Some(fname) = call.pointer("/function/name").and_then(|n| n.as_str()) {
495 *name = fname.to_string();
496 }
497 if let Some(args) = call.pointer("/function/arguments").and_then(|a| a.as_str())
498 {
499 let raw = self.raw_json.entry(idx).or_default();
500 raw.push_str(args);
501 let parsed: Value =
502 serde_json::from_str(raw).unwrap_or(Value::Object(Default::default()));
503 if let ContentBlock::ToolUse { input, .. } = &mut self.message.content[pos]
504 {
505 *input = parsed;
506 }
507 }
508 }
509 if event.is_none()
510 && let Some(ContentBlock::ToolUse { id, name, .. }) =
511 self.message.content.get(pos)
512 {
513 event = Some(StreamEvent::ToolUseStart {
514 index: pos,
515 id: id.clone(),
516 name: name.clone(),
517 partial: self.message.clone(),
518 });
519 }
520 }
521 }
522
523 Ok(event)
524 }
525}
526
527fn append_text_block(content: &mut Vec<ContentBlock>, text: &str) {
528 if let Some(ContentBlock::Text { text: last }) = content.last_mut() {
529 last.push_str(text);
530 return;
531 }
532 content.push(ContentBlock::text(text));
533}
534
535#[async_trait::async_trait]
536impl Provider for OpenAiCompletionsProvider {
537 fn name(&self) -> &str {
538 "openai-completions"
539 }
540
541 fn circuit_key(&self) -> String {
542 crate::gate::route_identity(self.name(), &self.config.base_url, &self.config.api_key)
543 }
544
545 async fn stream(
546 &self,
547 request: ChatRequest,
548 cancel: CancellationToken,
549 ) -> Result<EventStream, LlmError> {
550 let provider_permit = self.gate.acquire(&cancel).await?;
551 let url = format!(
552 "{}/chat/completions",
553 self.config.base_url.trim_end_matches('/')
554 );
555 let body = build_body(&self.config, &request)?;
556 let send_fut = self
557 .http
558 .post(&url)
559 .bearer_auth(&self.config.api_key)
560 .json(&body)
561 .send();
562 let response = tokio::select! {
563 _ = cancel.cancelled() => return Err(LlmError::Aborted { partial: None }),
564 r = send_fut => match r {
565 Ok(r) => r,
566 Err(e) => return Err(LlmError::Network(e.to_string())),
567 },
568 };
569
570 let status = response.status();
571 if !status.is_success() {
572 let retry_after = response
573 .headers()
574 .get("retry-after")
575 .and_then(|value| value.to_str().ok())
576 .and_then(|value| value.parse::<u64>().ok());
577 let text = response.text().await.unwrap_or_default();
578 return Err(map_status_error(status.as_u16(), &text, retry_after));
579 }
580
581 let model = request.model.clone();
582 let (mut sink, stream_rx) = channel(256);
583 let mut byte_stream = response.bytes_stream();
584 let mut decoder = SseDecoder::new();
585 let mut acc = Accumulator::new(&model);
586
587 tokio::spawn(async move {
588 loop {
589 tokio::select! {
590 _ = cancel.cancelled() => {
591 let partial = (!acc.message.content.is_empty()).then(|| Box::new(acc.message.clone()));
592 sink.close_error(LlmError::Aborted { partial }).await;
593 return;
594 }
595 chunk = byte_stream.next() => {
596 match chunk {
597 Some(Ok(bytes)) => {
598 decoder.push(&bytes);
599 while let Some(frame) = decoder.next_frame() {
600 let data = frame.data.trim();
601 if data == "[DONE]" {
602 // Content is ground truth here as
603 // everywhere else: endpoints that
604 // skip finish_reason entirely (seen
605 // on opencode-zen) still owe the loop
606 // a ToolUse signal when tool calls
607 // were streamed.
608 let has_tool_use = acc.message.content.iter().any(
609 |b| matches!(b, ContentBlock::ToolUse { .. }),
610 );
611 if has_tool_use {
612 acc.message.stop_reason = StopReason::ToolUse;
613 }
614 sink.close_message(acc.message.clone()).await;
615 return;
616 }
617 match acc.convert(data) {
618 Ok(Some(event)) => sink.push(event),
619 Ok(None) => {}
620 Err(e) => {
621 sink.close_error(e).await;
622 return;
623 }
624 }
625 }
626 }
627 Some(Err(e)) => {
628 sink.close_error(LlmError::Network(e.to_string())).await;
629 return;
630 }
631 None => {
632 // OpenAI-compatible proxies sometimes end the
633 // body after the last content chunk without
634 // [DONE]/finish_reason. A clean close with
635 // content is de facto completion (same as the
636 // [DONE] branch); empty content fails closed.
637 if acc.saw_end || !acc.message.content.is_empty() {
638 // Content is ground truth here too: a
639 // stream ending right after tool-call
640 // deltas must surface ToolUse, or the
641 // dangling call aborts the run.
642 let has_tool_use = acc.message.content.iter().any(
643 |b| matches!(b, ContentBlock::ToolUse { .. }),
644 );
645 if has_tool_use {
646 acc.message.stop_reason = StopReason::ToolUse;
647 }
648 sink.close_message(acc.message.clone()).await;
649 } else {
650 sink.close_error(LlmError::Parse(
651 "stream closed before finish_reason".into(),
652 )).await;
653 }
654 return;
655 }
656 }
657 }
658 }
659 }
660 });
661
662 Ok(stream_rx.with_guard(provider_permit))
663 }
664}
665
666#[cfg(test)]
667mod build_body_tests {
668 #![allow(clippy::unwrap_used, clippy::expect_used)]
669 use super::*;
670 use crate::types::CacheHints;
671
672 fn req_with_cache() -> ChatRequest {
673 let mut req = ChatRequest::new("gpt-5.6");
674 req.messages = vec![Message::user_text("hi")];
675 req.cache = Some(CacheHints {
676 session_key: "sess-1".into(),
677 breakpoints: Vec::new(),
678 });
679 req
680 }
681
682 #[test]
683 fn cache_key_off_sends_neither_hint() {
684 let config = OpenAiConfig::default();
685 let body = build_body(&config, &req_with_cache()).unwrap();
686 assert!(body.get("prompt_cache_key").is_none());
687 assert!(body.get("session_id").is_none());
688 }
689
690 #[test]
691 fn cache_key_on_sends_prompt_cache_key_only() {
692 let config = OpenAiConfig {
693 cache_key: true,
694 ..Default::default()
695 };
696 let body = build_body(&config, &req_with_cache()).unwrap();
697 assert_eq!(body["prompt_cache_key"], "sess-1");
698 assert!(body.get("session_id").is_none());
699 }
700
701 #[test]
702 fn openrouter_flag_adds_session_id_alongside_prompt_cache_key() {
703 let config = OpenAiConfig {
704 cache_key: true,
705 openrouter: true,
706 ..Default::default()
707 };
708 let body = build_body(&config, &req_with_cache()).unwrap();
709 assert_eq!(body["prompt_cache_key"], "sess-1");
710 assert_eq!(body["session_id"], "sess-1");
711 }
712
713 #[test]
714 fn no_cache_hint_on_request_sends_nothing_even_when_enabled() {
715 let config = OpenAiConfig {
716 api_key: String::new(),
717 base_url: String::new(),
718 cache_key: true,
719 openrouter: true,
720 };
721 let mut req = ChatRequest::new("gpt-5.6");
722 req.messages = vec![Message::user_text("hi")];
723 let body = build_body(&config, &req).unwrap();
724 assert!(body.get("prompt_cache_key").is_none());
725 assert!(body.get("session_id").is_none());
726 }
727}
728

1use futures::StreamExt;
2use serde_json::Value;
3use tokio_util::sync::CancellationToken;
4
5use crate::Provider;
6use crate::error::LlmError;
7use crate::gate::ProviderGate;
8use crate::sse::SseDecoder;
9use crate::stream::{EventStream, StreamEvent, channel};
10use crate::types::{
11 AssistantMessage, ChatRequest, ContentBlock, Message, Role, StopReason, ToolDefinition, Usage,
12};
13
14pub const OPENAI_DEFAULT_BASE_URL: &str = "https://api.openai.com/v1";
15
16#[derive(Debug, Clone, Default)]
17pub struct OpenAiConfig {
18 pub api_key: String,
19 pub base_url: String,
20 /// When true, and the request carries `ChatRequest::cache`, send
21 /// `prompt_cache_key` so the provider can route repeat traffic to the
22 /// same cache-warm backend.
23 pub cache_key: bool,
24 /// When true, also send OpenRouter's `session_id` alongside
25 /// `prompt_cache_key` — OpenRouter accepts both and uses `session_id`
26 /// to pin the upstream that holds the cache.
27 pub openrouter: bool,
28}
29
30/// Batch transcription through the OpenAI-compatible `/audio/transcriptions`
31/// endpoint. The model is supplied by discovery/configuration; this adapter
32/// deliberately has no baked-in model catalogue or default.
33pub async fn transcribe(
34 config: &OpenAiConfig,
35 audio: &[u8],
36 mime: &str,
37 model: &str,
38 cancel: &CancellationToken,
39) -> Result<String, LlmError> {
40 if audio.is_empty() || mime.trim().is_empty() || model.trim().is_empty() {
41 return Err(LlmError::InvalidRequest(
42 "audio, mime, and model are required".into(),
43 ));
44 }
45 if cancel.is_cancelled() {
46 return Err(LlmError::Aborted { partial: None });
47 }
48 let filename = if mime.contains("wav") {
49 "audio.wav"
50 } else if mime.contains("mpeg") || mime.contains("mp3") {
51 "audio.mp3"
52 } else if mime.contains("ogg") {
53 "audio.ogg"
54 } else if mime.contains("oga") {
55 "audio.oga"
56 } else if mime.contains("webm") {
57 "audio.webm"
58 } else if mime.contains("mp4") || mime.contains("m4a") {
59 "audio.m4a"
60 } else if mime.contains("flac") {
61 "audio.flac"
62 } else {
63 "audio.bin"
64 };
65 let part = reqwest::multipart::Part::bytes(audio.to_vec())
66 .file_name(filename)
67 .mime_str(mime)
68 .map_err(|e| LlmError::InvalidRequest(e.to_string()))?;
69 let form = reqwest::multipart::Form::new()
70 .part("file", part)
71 .text("model", model.to_string());
72 let url = format!(
73 "{}/audio/transcriptions",
74 config.base_url.trim_end_matches('/')
75 );
76 let response = tokio::select! {
77 _ = cancel.cancelled() => return Err(LlmError::Aborted { partial: None }),
78 result = reqwest::Client::new().post(url).bearer_auth(&config.api_key).multipart(form).send() => result.map_err(|e| LlmError::Network(e.to_string()))?,
79 };
80 let status = response.status().as_u16();
81 let value: Value = response
82 .json()
83 .await
84 .map_err(|e| LlmError::Parse(e.to_string()))?;
85 if status >= 400 {
86 return Err(LlmError::InvalidRequest(format!(
87 "transcription provider returned HTTP {status}: {value}"
88 )));
89 }
90 let text = value
91 .get("text")
92 .and_then(Value::as_str)
93 .unwrap_or("")
94 .trim()
95 .to_string();
96 Ok(text)
97}
98
99/// Synthesize speech through an OpenAI-compatible `/audio/speech` endpoint.
100/// The model is always explicit and the response is bounded before it is
101/// materialized, so a misbehaving provider cannot exhaust the process.
102pub async fn speak(
103 config: &OpenAiConfig,
104 text: &str,
105 model: &str,
106 voice: Option<&str>,
107 format: &str,
108 cancel: &CancellationToken,
109) -> Result<Vec<u8>, LlmError> {
110 if text.trim().is_empty() || model.trim().is_empty() || format.trim().is_empty() {
111 return Err(LlmError::InvalidRequest(
112 "text, model, and format are required".into(),
113 ));
114 }
115 if cancel.is_cancelled() {
116 return Err(LlmError::Aborted { partial: None });
117 }
118 let mut body = serde_json::json!({
119 "model": model,
120 "input": text,
121 "response_format": format,
122 });
123 if let Some(voice) = voice.filter(|v| !v.trim().is_empty()) {
124 body["voice"] = Value::String(voice.to_string());
125 }
126 let url = format!("{}/audio/speech", config.base_url.trim_end_matches('/'));
127 let response = tokio::select! {
128 _ = cancel.cancelled() => return Err(LlmError::Aborted { partial: None }),
129 result = reqwest::Client::new().post(url).bearer_auth(&config.api_key).json(&body).send() => result.map_err(|e| LlmError::Network(e.to_string()))?,
130 };
131 let status = response.status().as_u16();
132 if status >= 400 {
133 let body = response.text().await.unwrap_or_default();
134 return Err(map_status_error(status, &body, None));
135 }
136 const MAX_AUDIO_BYTES: usize = 16 * 1024 * 1024;
137 if response
138 .content_length()
139 .is_some_and(|n| n > MAX_AUDIO_BYTES as u64)
140 {
141 return Err(LlmError::InvalidRequest(
142 "provider audio exceeds 16 MiB".into(),
143 ));
144 }
145 let mut output = Vec::new();
146 let mut stream = response.bytes_stream();
147 while let Some(chunk) = tokio::select! {
148 _ = cancel.cancelled() => return Err(LlmError::Aborted { partial: None }),
149 chunk = stream.next() => chunk,
150 } {
151 let chunk = chunk.map_err(|e| LlmError::Network(e.to_string()))?;
152 if output.len().saturating_add(chunk.len()) > MAX_AUDIO_BYTES {
153 return Err(LlmError::InvalidRequest(
154 "provider audio exceeds 16 MiB".into(),
155 ));
156 }
157 output.extend_from_slice(&chunk);
158 }
159 if output.is_empty() {
160 return Err(LlmError::Parse("provider returned empty audio".into()));
161 }
162 Ok(output)
163}
164
165#[derive(Clone)]
166pub struct OpenAiCompletionsProvider {
167 http: reqwest::Client,
168 config: OpenAiConfig,
169 gate: ProviderGate,
170}
171
172impl OpenAiCompletionsProvider {
173 pub fn new(config: OpenAiConfig) -> Result<Self, LlmError> {
174 let http = reqwest::Client::builder()
175 .connect_timeout(std::time::Duration::from_secs(30))
176 .build()
177 .map_err(|e| LlmError::Network(e.to_string()))?;
178 Ok(OpenAiCompletionsProvider {
179 http,
180 gate: ProviderGate::new(&config.base_url, &config.api_key),
181 config,
182 })
183 }
184}
185
186pub fn build_body(config: &OpenAiConfig, request: &ChatRequest) -> Result<Value, LlmError> {
187 // Chat Completions carries no reasoning-item channel, so `Thinking`
188 // blocks are dropped unconditionally by `append_message` below — there
189 // is no turn boundary to compute here (contrast Anthropic/Google, which
190 // must replay signed thinking within the current turn).
191 let mut messages: Vec<Value> = Vec::with_capacity(request.messages.len() + 1);
192 if let Some(system) = &request.system {
193 messages.push(serde_json::json!({"role": "system", "content": system}));
194 }
195 for m in &request.messages {
196 append_message(&mut messages, m)?;
197 }
198
199 let mut body = serde_json::json!({
200 "model": request.model,
201 "messages": messages,
202 "stream": true,
203 "stream_options": {"include_usage": true},
204 });
205 if let Some(cache) = &request.cache {
206 if config.cache_key {
207 body["prompt_cache_key"] = serde_json::json!(cache.session_key);
208 }
209 if config.openrouter {
210 body["session_id"] = serde_json::json!(cache.session_key);
211 }
212 }
213 if !request.tools.is_empty() {
214 let tools: Vec<Value> = request
215 .tools
216 .iter()
217 .map(|t: &ToolDefinition| {
218 serde_json::json!({
219 "type": "function",
220 "function": {
221 "name": t.name,
222 "description": t.description,
223 "parameters": t.parameters,
224 },
225 })
226 })
227 .collect();
228 body["tools"] = Value::Array(tools);
229 }
230 Ok(body)
231}
232
233fn append_message(out: &mut Vec<Value>, m: &Message) -> Result<(), LlmError> {
234 match m.role {
235 Role::User => {
236 let mut text_parts: Vec<&str> = Vec::new();
237 let mut image_parts: Vec<String> = Vec::new();
238 let mut tool_results: Vec<&ContentBlock> = Vec::new();
239 for b in &m.content {
240 match b {
241 ContentBlock::Text { text } => text_parts.push(text),
242 ContentBlock::Image { source } => {
243 // Data URLs are the transport-agnostic form for the
244 // chat-completions API.
245 image_parts
246 .push(format!("data:{};base64,{}", source.media_type, source.data));
247 }
248 ContentBlock::ToolResult { .. } => tool_results.push(b),
249 ContentBlock::ToolUse { .. } => {
250 return Err(LlmError::InvalidRequest(
251 "tool_use blocks must appear in assistant messages".into(),
252 ));
253 }
254 ContentBlock::Thinking { .. } => {}
255 // Only the Anthropic adapter understands server-side
256 // tool search; every other adapter skips this opaque
257 // block entirely (docs/design/68 §5/§12).
258 ContentBlock::Provider { .. } => {}
259 }
260 }
261 for r in tool_results {
262 let ContentBlock::ToolResult {
263 tool_use_id,
264 content,
265 ..
266 } = r
267 else {
268 unreachable!()
269 };
270 out.push(serde_json::json!({
271 "role": "tool",
272 "tool_call_id": tool_use_id,
273 "content": content,
274 }));
275 }
276 if !image_parts.is_empty() || !text_parts.is_empty() {
277 let content = if image_parts.is_empty() {
278 serde_json::json!(text_parts.join("\n"))
279 } else {
280 let mut parts: Vec<Value> = text_parts
281 .iter()
282 .map(|t| serde_json::json!({"type": "text", "text": t}))
283 .collect();
284 for url in &image_parts {
285 parts.push(serde_json::json!({
286 "type": "image_url",
287 "image_url": {"url": url}
288 }));
289 }
290 serde_json::json!(parts)
291 };
292 out.push(serde_json::json!({
293 "role": "user",
294 "content": content,
295 }));
296 }
297 }
298 Role::Assistant => {
299 let mut text = String::new();
300 let mut tool_calls: Vec<Value> = Vec::new();
301 for b in &m.content {
302 match b {
303 ContentBlock::Text { text: t } => {
304 if !text.is_empty() {
305 text.push('\n');
306 }
307 text.push_str(t);
308 }
309 ContentBlock::ToolUse { id, name, input } => {
310 tool_calls.push(serde_json::json!({
311 "id": id,
312 "type": "function",
313 "function": {
314 "name": name,
315 "arguments": serde_json::to_string(input)
316 .map_err(|e| LlmError::Parse(e.to_string()))?,
317 },
318 }));
319 }
320 ContentBlock::Thinking { .. }
321 | ContentBlock::ToolResult { .. }
322 | ContentBlock::Image { .. }
323 | ContentBlock::Provider { .. } => {}
324 }
325 }
326 let mut msg = serde_json::json!({"role": "assistant"});
327 if !text.is_empty() || tool_calls.is_empty() {
328 msg["content"] = Value::String(text);
329 } else {
330 msg["content"] = Value::Null;
331 }
332 if !tool_calls.is_empty() {
333 msg["tool_calls"] = Value::Array(tool_calls);
334 }
335 out.push(msg);
336 }
337 }
338 Ok(())
339}
340
341fn map_status_error(status: u16, body: &str, retry_after: Option<u64>) -> LlmError {
342 let message = serde_json::from_str::<Value>(body)
343 .ok()
344 .and_then(|v| {
345 v.pointer("/error/message")
346 .or_else(|| v.pointer("/message"))
347 .and_then(|m| m.as_str().map(String::from))
348 })
349 .unwrap_or_else(|| body.chars().take(500).collect());
350
351 match status {
352 401 | 403 => LlmError::Auth(message),
353 400 => match LlmError::classify_400(message.clone()) {
354 over_length @ LlmError::Context(_) => over_length,
355 _ => LlmError::invalid_request_for_endpoint("/v1/chat/completions", message),
356 },
357 404 | 413 | 422 => LlmError::invalid_request_for_endpoint("/v1/chat/completions", message),
358 429 => LlmError::RateLimit {
359 message,
360 retry_after_secs: retry_after,
361 },
362 503 | 529 => LlmError::Overloaded(message),
363 _ => LlmError::Api { status, message },
364 }
365}
366
367struct Accumulator {
368 message: AssistantMessage,
369 saw_end: bool,
370 tool_pos: std::collections::HashMap<usize, usize>,
371 raw_json: std::collections::HashMap<usize, String>,
372}
373
374impl Accumulator {
375 fn new(model: &str) -> Self {
376 Accumulator {
377 message: AssistantMessage::empty(model),
378 saw_end: false,
379 tool_pos: std::collections::HashMap::new(),
380 raw_json: std::collections::HashMap::new(),
381 }
382 }
383
384 fn convert(&mut self, data: &str) -> Result<Option<StreamEvent>, LlmError> {
385 let v: Value = serde_json::from_str(data)
386 .map_err(|e| LlmError::Parse(format!("bad chunk json: {e}")))?;
387
388 if let Some(usage) = v.get("usage").filter(|u| !u.is_null()) {
389 // OpenAI's `prompt_tokens` is the WHOLE prompt, cache hits
390 // included (`prompt_tokens_details.cached_tokens` is a subset
391 // of it, not an addition to it). Normalized `input_tokens` is
392 // only the non-cached remainder, matching every other adapter
393 // (docs/design/68-context-engine.md §1) — `Usage::prompt_tokens()`
394 // reconstructs the original total.
395 let prompt_tokens = usage
396 .get("prompt_tokens")
397 .and_then(|x| x.as_u64())
398 .unwrap_or(0);
399 let cached_tokens = usage
400 .get("prompt_tokens_details")
401 .and_then(|d| d.get("cached_tokens"))
402 .and_then(|x| x.as_u64())
403 .unwrap_or(0);
404 self.message.usage = Usage {
405 input_tokens: prompt_tokens.saturating_sub(cached_tokens),
406 output_tokens: usage
407 .get("completion_tokens")
408 .and_then(|x| x.as_u64())
409 .unwrap_or(0),
410 cache_read_input_tokens: (cached_tokens > 0).then_some(cached_tokens),
411 cache_creation_input_tokens: None,
412 ..Default::default()
413 };
414 }
415
416 let Some(choice) = v.get("choices").and_then(|c| c.get(0)) else {
417 return Ok(None);
418 };
419
420 #[allow(clippy::dbg_macro)]
421 if std::env::var_os("VAK_LLM_DEBUG").is_some() {
422 eprintln!(
423 "[vak-llm] frame finish={:?} delta_keys={:?} content_len={}",
424 choice.get("finish_reason"),
425 choice.get("delta").map(|d| d
426 .as_object()
427 .map(|o| o.keys().cloned().collect::<Vec<_>>())
428 .unwrap_or_default()),
429 self.message.content.len()
430 );
431 }
432 if let Some(finish) = choice.get("finish_reason").and_then(|f| f.as_str()) {
433 // Accumulated tool_use blocks are ground truth: some compat
434 // endpoints close tool-call turns with finish reasons outside
435 // the canonical set (e.g. plain "stop"). Trusting the label
436 // would strand a dangling tool_use and kill the run.
437 let has_tool_use = self
438 .message
439 .content
440 .iter()
441 .any(|b| matches!(b, ContentBlock::ToolUse { .. }));
442 self.message.stop_reason = if has_tool_use {
443 StopReason::ToolUse
444 } else {
445 match finish {
446 "tool_calls" | "function_call" => StopReason::ToolUse,
447 "length" => StopReason::MaxTokens,
448 _ => StopReason::EndTurn,
449 }
450 };
451 self.saw_end = true;
452 return Ok(Some(StreamEvent::End {
453 message: self.message.clone(),
454 }));
455 }
456
457 let Some(delta) = choice.get("delta") else {
458 return Ok(None);
459 };
460
461 let mut event = None;
462 if let Some(text) = delta
463 .get("content")
464 .and_then(|c| c.as_str())
465 .filter(|t| !t.is_empty())
466 {
467 append_text_block(&mut self.message.content, text);
468 event = Some(StreamEvent::TextDelta {
469 delta: text.to_string(),
470 partial: self.message.clone(),
471 });
472 }
473
474 if let Some(calls) = delta.get("tool_calls").and_then(|c| c.as_array()) {
475 for call in calls {
476 let idx = call.get("index").and_then(|i| i.as_u64()).unwrap_or(0) as usize;
477 let pos = match self.tool_pos.get(&idx) {
478 Some(&p) => p,
479 None => {
480 self.message.content.push(ContentBlock::ToolUse {
481 id: String::new(),
482 name: String::new(),
483 input: Value::Object(Default::default()),
484 });
485 let p = self.message.content.len() - 1;
486 self.tool_pos.insert(idx, p);
487 p
488 }
489 };
490 if let ContentBlock::ToolUse { id, name, .. } = &mut self.message.content[pos] {
491 if let Some(new_id) = call.get("id").and_then(|i| i.as_str()) {
492 *id = new_id.to_string();
493 }
494 if let Some(fname) = call.pointer("/function/name").and_then(|n| n.as_str()) {
495 *name = fname.to_string();
496 }
497 if let Some(args) = call.pointer("/function/arguments").and_then(|a| a.as_str())
498 {
499 let raw = self.raw_json.entry(idx).or_default();
500 raw.push_str(args);
501 let parsed: Value =
502 serde_json::from_str(raw).unwrap_or(Value::Object(Default::default()));
503 if let ContentBlock::ToolUse { input, .. } = &mut self.message.content[pos]
504 {
505 *input = parsed;
506 }
507 }
508 }
509 if event.is_none()
510 && let Some(ContentBlock::ToolUse { id, name, .. }) =
511 self.message.content.get(pos)
512 {
513 event = Some(StreamEvent::ToolUseStart {
514 index: pos,
515 id: id.clone(),
516 name: name.clone(),
517 partial: self.message.clone(),
518 });
519 }
520 }
521 }
522
523 Ok(event)
524 }
525}
526
527fn append_text_block(content: &mut Vec<ContentBlock>, text: &str) {
528 if let Some(ContentBlock::Text { text: last }) = content.last_mut() {
529 last.push_str(text);
530 return;
531 }
532 content.push(ContentBlock::text(text));
533}
534
535#[async_trait::async_trait]
536impl Provider for OpenAiCompletionsProvider {
537 fn name(&self) -> &str {
538 "openai-completions"
539 }
540
541 fn circuit_key(&self) -> String {
542 crate::gate::route_identity(self.name(), &self.config.base_url, &self.config.api_key)
543 }
544
545 async fn stream(
546 &self,
547 request: ChatRequest,
548 cancel: CancellationToken,
549 ) -> Result<EventStream, LlmError> {
550 let provider_permit = self.gate.acquire(&cancel).await?;
551 let url = format!(
552 "{}/chat/completions",
553 self.config.base_url.trim_end_matches('/')
554 );
555 let body = build_body(&self.config, &request)?;
556 let send_fut = self
557 .http
558 .post(&url)
559 .bearer_auth(&self.config.api_key)
560 .json(&body)
561 .send();
562 let response = tokio::select! {
563 _ = cancel.cancelled() => return Err(LlmError::Aborted { partial: None }),
564 r = send_fut => match r {
565 Ok(r) => r,
566 Err(e) => return Err(LlmError::Network(e.to_string())),
567 },
568 };
569
570 let status = response.status();
571 if !status.is_success() {
572 let retry_after = response
573 .headers()
574 .get("retry-after")
575 .and_then(|value| value.to_str().ok())
576 .and_then(|value| value.parse::<u64>().ok());
577 let text = response.text().await.unwrap_or_default();
578 return Err(map_status_error(status.as_u16(), &text, retry_after));
579 }
580
581 let model = request.model.clone();
582 let (mut sink, stream_rx) = channel(256);
583 let mut byte_stream = response.bytes_stream();
584 let mut decoder = SseDecoder::new();
585 let mut acc = Accumulator::new(&model);
586
587 tokio::spawn(async move {
588 loop {
589 tokio::select! {
590 _ = cancel.cancelled() => {
591 let partial = (!acc.message.content.is_empty()).then(|| Box::new(acc.message.clone()));
592 sink.close_error(LlmError::Aborted { partial }).await;
593 return;
594 }
595 chunk = byte_stream.next() => {
596 match chunk {
597 Some(Ok(bytes)) => {
598 decoder.push(&bytes);
599 while let Some(frame) = decoder.next_frame() {
600 let data = frame.data.trim();
601 if data == "[DONE]" {
602 // Content is ground truth here as
603 // everywhere else: endpoints that
604 // skip finish_reason entirely (seen
605 // on opencode-zen) still owe the loop
606 // a ToolUse signal when tool calls
607 // were streamed.
608 let has_tool_use = acc.message.content.iter().any(
609 |b| matches!(b, ContentBlock::ToolUse { .. }),
610 );
611 if has_tool_use {
612 acc.message.stop_reason = StopReason::ToolUse;
613 }
614 sink.close_message(acc.message.clone()).await;
615 return;
616 }
617 match acc.convert(data) {
618 Ok(Some(event)) => sink.push(event),
619 Ok(None) => {}
620 Err(e) => {
621 sink.close_error(e).await;
622 return;
623 }
624 }
625 }
626 }
627 Some(Err(e)) => {
628 sink.close_error(LlmError::Network(e.to_string())).await;
629 return;
630 }
631 None => {
632 // OpenAI-compatible proxies sometimes end the
633 // body after the last content chunk without
634 // [DONE]/finish_reason. A clean close with
635 // content is de facto completion (same as the
636 // [DONE] branch); empty content fails closed.
637 if acc.saw_end || !acc.message.content.is_empty() {
638 // Content is ground truth here too: a
639 // stream ending right after tool-call
640 // deltas must surface ToolUse, or the
641 // dangling call aborts the run.
642 let has_tool_use = acc.message.content.iter().any(
643 |b| matches!(b, ContentBlock::ToolUse { .. }),
644 );
645 if has_tool_use {
646 acc.message.stop_reason = StopReason::ToolUse;
647 }
648 sink.close_message(acc.message.clone()).await;
649 } else {
650 sink.close_error(LlmError::Parse(
651 "stream closed before finish_reason".into(),
652 )).await;
653 }
654 return;
655 }
656 }
657 }
658 }
659 }
660 });
661
662 Ok(stream_rx.with_guard(provider_permit))
663 }
664}
665
666#[cfg(test)]
667mod build_body_tests {
668 #![allow(clippy::unwrap_used, clippy::expect_used)]
669 use super::*;
670 use crate::types::CacheHints;
671
672 fn req_with_cache() -> ChatRequest {
673 let mut req = ChatRequest::new("gpt-5.6");
674 req.messages = vec![Message::user_text("hi")];
675 req.cache = Some(CacheHints {
676 session_key: "sess-1".into(),
677 breakpoints: Vec::new(),
678 });
679 req
680 }
681
682 #[test]
683 fn cache_key_off_sends_neither_hint() {
684 let config = OpenAiConfig::default();
685 let body = build_body(&config, &req_with_cache()).unwrap();
686 assert!(body.get("prompt_cache_key").is_none());
687 assert!(body.get("session_id").is_none());
688 }
689
690 #[test]
691 fn cache_key_on_sends_prompt_cache_key_only() {
692 let config = OpenAiConfig {
693 cache_key: true,
694 ..Default::default()
695 };
696 let body = build_body(&config, &req_with_cache()).unwrap();
697 assert_eq!(body["prompt_cache_key"], "sess-1");
698 assert!(body.get("session_id").is_none());
699 }
700
701 #[test]
702 fn openrouter_flag_adds_session_id_alongside_prompt_cache_key() {
703 let config = OpenAiConfig {
704 cache_key: true,
705 openrouter: true,
706 ..Default::default()
707 };
708 let body = build_body(&config, &req_with_cache()).unwrap();
709 assert_eq!(body["prompt_cache_key"], "sess-1");
710 assert_eq!(body["session_id"], "sess-1");
711 }
712
713 #[test]
714 fn no_cache_hint_on_request_sends_nothing_even_when_enabled() {
715 let config = OpenAiConfig {
716 api_key: String::new(),
717 base_url: String::new(),
718 cache_key: true,
719 openrouter: true,
720 };
721 let mut req = ChatRequest::new("gpt-5.6");
722 req.messages = vec![Message::user_text("hi")];
723 let body = build_body(&config, &req).unwrap();
724 assert!(body.get("prompt_cache_key").is_none());
725 assert!(body.get("session_id").is_none());
726 }
727}
728

Openai

Indexing the workspace…

Openai